diff --git a/src/cpu/o3/lsq_unit.hh b/src/cpu/o3/lsq_unit.hh
index d964b9f9f..be9224099 100644
--- a/src/cpu/o3/lsq_unit.hh
+++ b/src/cpu/o3/lsq_unit.hh
@@ -643,7 +643,10 @@ LSQUnit<Impl>::read(Request *req, T &data, int load_idx)
     // if we the cache is not blocked, do cache access
     if (!lsq->cacheBlocked()) {
         PacketPtr data_pkt =
-            new Packet(req, MemCmd::ReadReq, Packet::Broadcast);
+            new Packet(req,
+                       (req->isLocked() ?
+                        MemCmd::LoadLockedReq : MemCmd::ReadReq),
+                       Packet::Broadcast);
         data_pkt->dataStatic(load_inst->memData);
 
         LSQSenderState *state = new LSQSenderState;
diff --git a/src/cpu/o3/lsq_unit_impl.hh b/src/cpu/o3/lsq_unit_impl.hh
index 810a6d29f..5ae1cc0e4 100644
--- a/src/cpu/o3/lsq_unit_impl.hh
+++ b/src/cpu/o3/lsq_unit_impl.hh
@@ -647,7 +647,9 @@ LSQUnit<Impl>::writebackStores()
 
         memcpy(inst->memData, storeQueue[storeWBIdx].data, req->getSize());
 
-        MemCmd command = req->isSwap() ? MemCmd::SwapReq : MemCmd::WriteReq;
+        MemCmd command =
+            req->isSwap() ? MemCmd::SwapReq :
+            (req->isLocked() ? MemCmd::WriteReq : MemCmd::StoreCondReq);
         PacketPtr data_pkt = new Packet(req, command,
                                         Packet::Broadcast);
         data_pkt->dataStatic(inst->memData);
diff --git a/src/cpu/ozone/lw_lsq.hh b/src/cpu/ozone/lw_lsq.hh
index d9e0d04ac..ba40e9ce1 100644
--- a/src/cpu/ozone/lw_lsq.hh
+++ b/src/cpu/ozone/lw_lsq.hh
@@ -632,7 +632,11 @@ OzoneLWLSQ<Impl>::read(RequestPtr req, T &data, int load_idx)
     DPRINTF(OzoneLSQ, "Doing timing access for inst PC %#x\n",
             inst->readPC());
 
-    PacketPtr data_pkt = new Packet(req, Packet::ReadReq, Packet::Broadcast);
+    PacketPtr data_pkt =
+        new Packet(req,
+                   (req->isLocked() ?
+                    MemCmd::LoadLockedReq : Packet::ReadReq),
+                   Packet::Broadcast);
     data_pkt->dataStatic(inst->memData);
 
     LSQSenderState *state = new LSQSenderState;
diff --git a/src/cpu/ozone/lw_lsq_impl.hh b/src/cpu/ozone/lw_lsq_impl.hh
index e3000288c..82191312a 100644
--- a/src/cpu/ozone/lw_lsq_impl.hh
+++ b/src/cpu/ozone/lw_lsq_impl.hh
@@ -587,7 +587,10 @@ OzoneLWLSQ<Impl>::writebackStores()
         memcpy(inst->memData, (uint8_t *)&(*sq_it).data,
                req->getSize());
 
-        PacketPtr data_pkt = new Packet(req, Packet::WriteReq, Packet::Broadcast);
+        MemCmd command =
+            req->isSwap() ? MemCmd::SwapReq :
+            (req->isLocked() ? MemCmd::WriteReq : MemCmd::StoreCondReq);
+        PacketPtr data_pkt = new Packet(req, command, Packet::Broadcast);
         data_pkt->dataStatic(inst->memData);
 
         LSQSenderState *state = new LSQSenderState;
diff --git a/src/cpu/simple/atomic.cc b/src/cpu/simple/atomic.cc
index 8e8da2fa2..01eb4873e 100644
--- a/src/cpu/simple/atomic.cc
+++ b/src/cpu/simple/atomic.cc
@@ -280,7 +280,10 @@ AtomicSimpleCPU::read(Addr addr, T &data, unsigned flags)
 
     // Now do the access.
     if (fault == NoFault) {
-        Packet pkt = Packet(req, MemCmd::ReadReq, Packet::Broadcast);
+        Packet pkt =
+            Packet(req,
+                   req->isLocked() ? MemCmd::LoadLockedReq : MemCmd::ReadReq,
+                   Packet::Broadcast);
         pkt.dataStatic(&data);
 
         if (req->isMmapedIpr())
@@ -370,23 +373,24 @@ AtomicSimpleCPU::write(T data, Addr addr, unsigned flags, uint64_t *res)
 
     // Now do the access.
     if (fault == NoFault) {
-        Packet pkt =
-            Packet(req, req->isSwap() ? MemCmd::SwapReq : MemCmd::WriteReq,
-                   Packet::Broadcast);
-        pkt.dataStatic(&data);
-
+        MemCmd cmd = MemCmd::WriteReq; // default
         bool do_access = true;  // flag to suppress cache access
 
         if (req->isLocked()) {
+            cmd = MemCmd::StoreCondReq;
             do_access = TheISA::handleLockedWrite(thread, req);
+        } else if (req->isSwap()) {
+            cmd = MemCmd::SwapReq;
+            if (req->isCondSwap()) {
+                assert(res);
+                req->setExtraData(*res);
+            }
         }
-        if (req->isCondSwap()) {
-             assert(res);
-             req->setExtraData(*res);
-        }
-
 
         if (do_access) {
+            Packet pkt = Packet(req, cmd, Packet::Broadcast);
+            pkt.dataStatic(&data);
+
             if (req->isMmapedIpr()) {
                 dcache_latency = TheISA::handleIprWrite(thread->getTC(), &pkt);
             } else {
@@ -395,12 +399,14 @@ AtomicSimpleCPU::write(T data, Addr addr, unsigned flags, uint64_t *res)
             }
             dcache_access = true;
             assert(!pkt.isError());
+
+            if (req->isSwap()) {
+                assert(res);
+                *res = pkt.get<T>();
+            }
         }
 
-        if (req->isSwap()) {
-            assert(res);
-            *res = pkt.get<T>();
-        } else if (res) {
+        if (res && !req->isSwap()) {
             *res = req->getExtraData();
         }
     }
diff --git a/src/cpu/simple/timing.cc b/src/cpu/simple/timing.cc
index b4e4a4433..77df2c05d 100644
--- a/src/cpu/simple/timing.cc
+++ b/src/cpu/simple/timing.cc
@@ -260,7 +260,10 @@ TimingSimpleCPU::read(Addr addr, T &data, unsigned flags)
     // Now do the access.
     if (fault == NoFault) {
         PacketPtr pkt =
-            new Packet(req, MemCmd::ReadReq, Packet::Broadcast);
+            new Packet(req,
+                       (req->isLocked() ?
+                        MemCmd::LoadLockedReq : MemCmd::ReadReq),
+                       Packet::Broadcast);
         pkt->dataDynamic<T>(new T);
 
         if (!dcachePort.sendTiming(pkt)) {
@@ -350,25 +353,27 @@ TimingSimpleCPU::write(T data, Addr addr, unsigned flags, uint64_t *res)
 
     // Now do the access.
     if (fault == NoFault) {
-        assert(dcache_pkt == NULL);
-        if (req->isSwap())
-            dcache_pkt = new Packet(req, MemCmd::SwapReq, Packet::Broadcast);
-        else
-            dcache_pkt = new Packet(req, MemCmd::WriteReq, Packet::Broadcast);
-        dcache_pkt->allocate();
-        dcache_pkt->set(data);
-
+        MemCmd cmd = MemCmd::WriteReq; // default
         bool do_access = true;  // flag to suppress cache access
 
+        assert(dcache_pkt == NULL);
+
         if (req->isLocked()) {
+            cmd = MemCmd::StoreCondReq;
             do_access = TheISA::handleLockedWrite(thread, req);
-        }
-        if (req->isCondSwap()) {
-             assert(res);
-             req->setExtraData(*res);
+        } else if (req->isSwap()) {
+            cmd = MemCmd::SwapReq;
+            if (req->isCondSwap()) {
+                assert(res);
+                req->setExtraData(*res);
+            }
         }
 
         if (do_access) {
+            dcache_pkt = new Packet(req, MemCmd::WriteReq, Packet::Broadcast);
+            dcache_pkt->allocate();
+            dcache_pkt->set(data);
+
             if (!dcachePort.sendTiming(dcache_pkt)) {
                 _status = DcacheRetry;
             } else {
@@ -609,7 +614,7 @@ TimingSimpleCPU::completeDataAccess(PacketPtr pkt)
 
     Fault fault = curStaticInst->completeAcc(pkt, this, traceData);
 
-    if (pkt->isRead() && pkt->req->isLocked()) {
+    if (pkt->isRead() && pkt->isLocked()) {
         TheISA::handleLockedRead(thread, pkt->req);
     }